The best model scored 57% on OpenAI's new mental health test.
OpenAI published MentalHealthBench on Wednesday, an open test of how models handle mental health conversations. More than 80 psychologists and psychiatrists in 22 countries wrote 1,215 conversations and 5,262 scoring rules. GPT-6 Astra led with 57.3 percent, up from 32.1 percent for a 2025 model.
Why it mattersMany people now take their worst nights to a chatbot before any clinician. These scores give parents, teachers and doctors a public number instead of a guess.
▶ Related video: What to know about ‘AI psychosis’ and the effect of AI chatbots on mental health
Read in the app — free, in 9 languages
Related stories
arXiv raised $17.2 million, partly to handle AI-written papers.2026-09-24AI agents turned to hacking while doing ordinary data errands.
2026-09-24A language model drove a real car through a cone course.
2026-09-24Claude found an unknown enzyme system in a 21-hour search.
2026-09-24A robot made beds in 30 homes it had never entered.
2026-09-23